i'm doing lora extended
i'm doing lora extended
AWas it overtrained? Usually the LR for Lion is 1e-7
Zi don't think so, typically overtrained it looks burnt I thought.
Zi've been trying to train a model of my wife for ages now, and I get close, but never really there. I've followed every video I've come across, but still can't get it perfect
Yhelp !! this error on dreambooth 1111 i'ts new (again ) PermissionError: [WinError 5] Accès refusé: 'C:\Users\UTILIS~1\AppData\Local\Temp\tmp8_yihns4\text_encoder\model.safetensors'
PI'm unable to have both python 3.10.6 and 3.10.9, when I try to install 3.10.9 it is asking to upgrade from 3.10.6
AGitHub
pyenv for Windows. pyenv is a simple python version management tool. It lets you easily switch between multiple versions of Python. It's simple, unobtrusive, and follows the UNIX tradition ...
Fvery easy. each one have their own venv
Fi think they improve significantly because i train on 768x1024
Fi did a training for @eggbom
Fand you cant believe what raw quality we get
Fwith just
Fphoto of ohwx man
Feven no negatives
Fyes but they may bleed. likely to bleed if they are same class
Fnot yet. too many stuff i have to process :d
Fhi welcome
Fdelete that file and restart
P@Dr. Furkan GözükaraI have noticed that the quality of training on kohya is worse recently, is it the same for you?
Fi am using older commit :d
Fi also updated readme file
Fshared in my video description
Pthe latest commit works for me
Fworks but dont learn
Fno errors :d
Pit learns!
Fweird didnt learn for me
Fi even opened a topic
Fbut i did a fresh install
Fdidnt use older venv
Pdid you use torch 2 or 1?
Fi think torch 1
Ptorch 2 worked for me, using rtx 3090
Fi should retest
Pif you could, try to use like a one week old commit and compare it to the most recent commit
Fi used my video commit
Palso this is not just local machine but even runpod also worse training quality
Yok i'm deleting it and i figure how to upgrade khoya and create a dreambooth with it but it create only one model and i seek a way to doing like automatic 1111 : multiple models so i can fiond wich one is not too much trained
AInteresting LR for Adan Dadaptation Optimizer
Fyou use buckets? or train for that resolution?
AVRAM usage is a bit hectic
Fthat's a lot lol ,worth it?
AProbably not, I am still training, it/sec is almost 1/3 of using Lion, so very slow
Nbut if they are different concept, then it doesnt really matter if you did 2 concepts at once or just one training session after another right
Fno i use single bucket. all images are 768x1024
Fit matters since you are also training unet
Fthe more you train unet = bad
NI see so in multiple concepts, it doesnt train the unet so theres some advantage
Fso you train that resolution? SD 1.5 model?
Frealistic vision yes